DEVELOPMENT OF SUPERVISED AND UNSUPERVISED MACHINE LEARNING ALGORITHM FOR DETECTION OF MALARIA PARASITES IN THIN BLOOD SMEARS USING ORANGE SOFTWARE

Abstract

ABSTRACT

Highly sensitive and specific malaria diagnosis methods that are satisfactory for point-of-care testing in high burden areas are essential for productive treatment and monitoring of the disease. Microscopists often examine thick and thin blood smears which are the gold standard to diagnose malaria disease and compute parasitemia, Hence, the need for highly trained experts to interpret the data. In this study, machine learning algorithms for the detection of malaria parasite in thin blood smear images have been developed to reduce reliance on human proficiency, especially in the situations where experts are unavailable. The datasets containing 27558 cell images was obtained from National Library of Medicine, National Institute of Health (NIH) and used for both supervised and unsupervised machine learning models development. For supervised learning, logistic regression and random forest classifiers were used to predict the classes of thin blood smear images. These models classified the images as either uninfected or parasitised. Logistic regression returned a classification accuracy of 93.5% for parasitised images and 96.5% for uninfected smears. Random forest returned a classification accuracy of 90.5% for parasitised and 90.4% for uninfected smears. For unsupervised machine learning, hierarchical clustering and k-means models were implemented. Hierarchical clustering grouped parasitised images in one cluster and uninfected in another cluster and k-means gave a value of 0.218, discovered two clusters from the dataset. These results showed that logistic regression model produced the best performance for classification of thin blood smears of malaria. In cases where the classes of the smears are not known, the unsupervised machine learning models can be used to detect malaria infections in the smears. These models can be combined as backend programs for the design of a robust computerised malaria detection computer program. It is important to note that, although this method may not fully abolish the need for trained experts, the model implementations can be of great assistance in aiding the diagnostic decision-making process.